How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
In a Nutshell
Anthropic's product team ships features in days or weeks by removing all barriers, launching in research previews for rapid feedback, empowering engineers with product taste, and using clear goals, team principles, and tight cross-functional processes to align on a safe AGI mission. PM roles shift from long-term roadmaps to eliciting max capability from current models via first-principles thinking, evals, and blurring lines with engineering—prioritizing "product taste" over traditional skills. Key tools like Claude Code (CLI/desktop for coding) and Co-work (non-code tasks, decks, meeting prep) accelerate PM workflows, with advice to automate repetitive tasks to 100% success, stay "AGI pilled" realistically, and embrace chaos for fast iteration.
These notes were generated by AI and may contain inaccuracies.
It's very hard to be the right amount of AGI pilled. It's very easy to build the product for the super AGI strong model. The hard thing is figuring out for the current model, how do you elicit the maximum capability?
Anthropic ships at an unprecedented pace. Timelines for product features have gone down from 6 months to 1 month and sometimes to even one day. They want to remove every single barrier to shipping things.
PMs approaching it incorrectly by focusing on traditional methods. The PM role is changing quickly. For AI native products, iterating quickly and launching features every single week is extremely important.
Emerging skills for PMs come back to product taste. As code becomes much cheaper to write, deciding what to write becomes more valuable.
Cat Wu is Head of Product for Claude Code and co-work at Anthropic. Her team is building products changing how we build products.
Cat works alongside Boris, who created Claude Code, leads the team, and ships a bazillion PRs a day from his phone. Boris is the product visionary, setting the vision for what the product needs to be in 3-6 months, the AGI pilled version.
Cat's role is figuring out the path from today to that vision. She spends more time on cross-functional alignment: ensuring marketing, sales, finance, capacity are bought in, all rowing in the same direction, no blockers to shipping.
They mindmeld 80%, with blurry lines. 20% of things Cat drives where she cares more, 20% where Boris drives.
Cat interviews hundreds of PMs and sees them approaching it wrong. Before AI, tech shifts were slower, planning on 6-12 month horizons. Shipping slower meant emphasis on coordinating multi-quarter roadmaps with partner teams because code was expensive.
Now with AI accelerating engineering and model capabilities improving quickly, feature timelines are 1 month, 1 week, or 1 day. PMs need less emphasis on aligning roadmaps, more on fastest way to get something out the door.
Top PMs shorten time from idea to product in users' hands, define most important tasks that need to work out of the box.
-
Set clear goals because LLMs are general, creating ambiguity in users, problems, use cases. Example: Key user is professional developers. Problem: too many permission prompts causing fatigue. Use case: professional developers at enterprises safely get to zero permission prompts. This rules out approaches, enables one prompt to get more done.
-
Repeatable process for shipping: Ship almost all features in research preview, clearly branded as early product for feedback, may not be supported forever. Reduces commitment, ships in 1-2 weeks.
-
Framework for cross-functional: Tight process between engineering, marketing, docs. Engineers post ready features (after internal dogfooding) in evergreen launch room. Sarah (docs lead), Alex (PMM lead), Tar, Lydia (DevRel) turn around marketing announcement next day. Lowers friction for engineers to ship. PM sets this up.
Rigorous weekly metrics readouts with entire team to ensure deep understanding of business facets, key goals, trends, drivers.
List of team principles: key users, why those users, what we're willing to trade off. Enables autonomous decisions without blocking on PM or stakeholders.
PRDs still used for ambiguous features: 1-pager on goals, delightful use cases, failure modes to fix. Also for heavy infrastructure projects taking months.
Anthropic ships major features or products every day. Not primarily due to Mythos (in preview, powerful). They've moved fast for several quarters using models internally, but bulk from process and expectations.
Low on process, remove every barrier. Empower everyone to take idea to world in <1 week, sometimes 1 day.
Advantage of frontier models accelerates building and using products.
Claude Code source code leaked due to human error: human working with Claude to write PR updating package releases, passed two layers of human review. Hardened processes with more safeguards, most shipped. Person still at Anthropic; it's a process failure, focus on learning.
Open Claude: High demand for Claude, scaling infrastructure, making harness token efficient. Not designed for third-party products with different usage patterns. Prioritize first-party products and API. Offer credits with subscriptions for seamless transition.
Subsidizing at $200/month essentially unlimited use; can't give away compute when in demand for profitability.
~30-40 PMs in teams:
- Research PM team (Diane leads): Understand customer feedback for models, feed to research team, shepherd model launches.
- Claude Developer Platform team: Maintains APIs Claude Code built on, releases like managed agents (build agents, host on behalf).
- Claude Code team: Works on Claude Code and co-work core products.
- Enterprise team: Cost controls, RBAC, security for enterprise adoption.
- Growth team: Grows product suite, works closely on Claude Code, co-work, Claude API (CDP) growth.
Roles merging: PMs do engineering, engineers do PM, designers PM and land code.
Options: Hire more engineers with great product taste, or same engineers + more PMs to guide.
Focus on engineers with product taste: End-to-end from Twitter user feedback to shipping in a week with minimal product involvement. Most efficient.
Engineer/PM overlapping; benefit from more of either. Product taste rare; hire anyone demonstrating it strongly.
Cat's background: Engineer for years, brief VC, then Anthropic. Almost all PMs engineers or ship code on Claude Code. Designers were front-end engineers. Builds trust, moves faster.
Product taste most important: As code cheaper, deciding what to write, right UX, delightful experience. Tens of thousands GitHub issues; taste to prioritize which to build and how.
Engineering background useful short-term (next few months) for sensing build difficulty, aiding prioritization (easy: spend hour; hard: costs more).
Valued skill sets change frequently, hard to predict >few months. Large coding capability increases every few months change roles.
First principles thinking: Understand tech landscape changes, team needs, fill gaps, learn skills or apply existing. Work amorphous; great PMs identify highest priority gaps, low ego, wear/swap hats to move team faster.
Picking what to work on, knowing market direction, prioritizing, knowing if built thing is good/right, getting early version out.
Humans provide common sense models lack: thousand moving pieces in launches, small things that go wrong. Sense of stakeholders, relations, preferences, communication venues.
Tacit common sense, EQ valuable now; models will improve but gaps remain.
Team leans into chaos, faces challenges with smile. P0 issues escalate rapidly (Sunday P0 → Monday P0 → P0000). Acknowledge limits, sleep for good decisions, brutally prioritize, okay letting things go.
Ship products not as polished as wished. Look for people excited by hard challenges, do best, sleep knowing they did.
Top goal is to empower professional developers. If a product isn't successful but not blocking core use case, it's okay—hear feedback and fix in next release. Launching buggy features used to keep speaker up at night, but now acceptable with quick feedback loop and fixes in next release.
Analogy to Pirates of the Caribbean scene: guy strolling down stairs on ship as everything demolishes around him. Anthropic team is chill and optimistic; without it, burnout happens. Hire experienced people who have seen ups and downs, know what gives them energy, and maintain it over time.
Roles blurring: engineers becoming PMs. Potential losses: career ladders, clear paths, design consistency, code quality.
Sacrificing product consistency. Historically, code expensive, so planned products carefully: relations, use cases, integrations, one product per use case. Now with fast AI and ideas to test, features overlap—often testing two form factors internally to see which external users prefer.
For new users: harder to know best path for task X. Need more education on core features and best practices. Cost of launching many features.
Users feel hard to keep up; traditional PM ships monthly/quarterly, easy to check in occasionally without FOMO. With agentic tools (Claude Code, Co-work, ecosystem), people check Twitter daily for latest.
Goal: tools educate users, teach what they need, make them feel along for ride, less on fast treadmill.
/powerup feature walks through cool ways and best practices for Claude Code.
Past: avoided onboarding, product should be intuitive. Now, too many features, demand for built-in onboarding. Diverged from no-onboarding principle; users want top 10 of 100 features.
Anthropic behind at start: least funded, no distribution, not first, OpenAI ahead. Now killing it, $11B ARR, massive growth.
Two most important: 1) unifying mission—bring safe AGI to humanity. Hire people who care most. Reference in decisions on product focus. Put mission above individual products for fast, unified execution across org—unseen at this scale.
Clear mission (safety alignment, AI good for world) makes decisions easier: competing priorities resolved by mission importance, everyone aligns, even deprioritizing own work.
Explains focus vs. others (e.g., not launching social network or feed). Mission: put Anthropic goals ahead of individual/team goals/KRs. Teams sacrifice for company goals happily. Extreme: if Claude Code failed but Anthropic succeeded, extremely happy.
OpenClaude decision: prioritize growing users via Claude subscriptions/first-party products, at expense of third-party sometimes.
Claude Code: terminal/CLI for one-off coding tasks, latest features first (most powerful). Kick off handful of tasks.
Desktop: shines for front-end work, preview feature (e.g., web app: preview pane sees real-time changes while chatting). Graphical, less scary than terminal popups. At-a-glance view of all tasks: CLI sessions, other desktop, web/mobile. One-stop control plane.
Web/mobile: kick off tasks on the go (CLI/desktop need local laptop). Solves tethered laptop outside issue.
Co-work: non-code outputs—Slack/inbox zero, slide decks, docs on feature goals/launch plans.
Split: code output → Claude Code/Desktop/mobile; non-code → Co-work.
Co-work growing fast, underappreciated.
Connect data sources first: Google Calendar, Slack, Gmail, Google Drive for context (pulls threads, improves quality).
Example: Code with Claude conference talks. For agent transition talk, showcase products shipped, internal success stories/demos. Fed Google Drive, Slack, PMM Alex's draft points. Told narrative. Co-work worked 1 hour: checked Twitter launches, evergreen launch room, Claude Code announce channel (team demos). Synthesized 20-page deck overnight—pretty good, minimal tweaks (less wordy). Faster than manual; used design system, looks Anthropic-polished.
Prompt: "Make slide deck for Code with Claude conference. PMM suggested coverage. Current/manual drafts I don't like [links]. Create proposed outline with details, avoid overlap with keynote."
Claude proposed outline/ideas; PM decided final (e.g., progression: local tasks → every PR green → land more PRs, best demos). Co-work built deck in hours.
PM role: Claude brainstorms/synthesizes fast, presents options; PM decides final.
Design system: gave access to standardized external deck (colors, fonts, 20 slide formats). Or connect Figma (slide formats).
Stack: heavily Claude Code, Co-work. Anthropic runs on Slack (core OS, 30% time pushing Co-work boundaries, talking to model on mistakes).
Claude Code lowers barrier for custom apps/personalized work software vs. ill-fitting tools.
Sales example: repetitive decks (101, 201, mastering Claude Code). Web app inputs customer context from Salesforce, Gong, notes. Customizes: customer stack (Bedrock/Claude Enterprise/Console), concerns (code review), compliance (HIPAA/security). Adds/removes slides. 20-30 min manual → seconds, tailored deck.
Slack: durable comms infrastructure, real-time updates. Cutting-edge teams hooked. Hackable (Slack bots).
Engineering biggest token spender.
Applied AI second: push boundaries of Claude Code/Co-work. Help customers adopt API; make prototypes (faster with Claude Code). Manage customer comms/inbound/historical notes (heavy Co-work/Code).
Applied AI: technical go-to-market, help adopt latest API/models for customer products/internal acceleration. Like forward-deployed engineering + customer success.
They cover multiple customers (5-10 engagements/day). Night before: Co-work summarizes upcoming customer meetings.
The night before customer meetings, PI persons ask Co-work to summarize upcoming meetings, customer requests, top-of-mind issues, and action items from past meetings. Co-work compiles a dossier or brief for awareness going into the meeting. Co-work can research answers, such as feature launch ETAs by searching Slack, adding the latest info to notes for use during calls. These are workflows people build and share with their team.
Token spend exceeds some people's salaries in AI usage. At Anthropic, as models improve, people delegate more tasks, spending more hours in tools like Claude Code and Co-work. Token cost per engineer or knowledge worker increases with each model jump or product improvement. It remains lower than average engineer salary, but the percentage is increasing.
Anthropic provides a lot of tokens, though some hit limits. They empower teams to build fast, trusting understanding of serving costs. Wasting tokens is frowned upon, but individuals judge responsibly.
Hardest skill for PMs: defining product vision one month out amid ambiguity in model capabilities and user behavior. Best PMs spot patterns from users abusing existing product limits, set direction, execute steadily, and adjust if model capabilities differ from expectations.
Hard to be the right amount of AGI pilled. Super AGI future means simple text box suffices—model adds tools, integrations, asks questions when uncertain. Hard part: eliciting maximum capability from current models, guiding users to golden path, leveraging strengths, patching weaknesses. This skill is rare.
Building the Skill
Spend time talking to and using the model. Ask model to introspect on behaviors, e.g., why it made front-end changes but skipped UI tests. Model might reveal confusing system prompt, overlooked verification, or unchecked sub-agent work. Curiosity about model decisions reveals fixes for harness gaps.
Identify trusted users (handful of 5) who articulate model/harness quality accurately—not all feedback is qualified.
Build evals: 10 great evals quantify goals, progress, gaps. Underappreciated; more PMs/engineers should do it.
Eval importance varies by feature. Some team members collaborate with research on Claude Code behaviors. Cat jumps into evals for features needing definition, e.g., outputs 5 evals, success rates, prompts to improve.
Features like memory benefit. Human evals from skilled people spot spikes/lacks.
Shoutouts for Model Evaluation
- Amanda: Molds Claude's character—ambiguous task harder than coding. Articulates goals, success/failure with conviction.
- Claude Code team: Team lunches for new model vibes, e.g., not explaining thinking, abrupt, low-quality memories, testing habits. Informs data hypotheses.
Claude's character is core to success. Users love Claude Code/Co-work's lighthearted, fun, competent vibe; low ego (genuinely sorry, fixes collaboratively); positive (breaks down tasks, offers to start). Great co-worker traits: positivity, bias to action, earnest feedback. Imbued to make work enjoyable.
New models often require revisiting: remove unneeded features added as model crutches.
Example: To-do list for Claude Code refactors. Early models changed 5/20 call sites then stopped. Added to-do list mimicking VS Code (list call sites, check off). Improved completion. Later models (Opus 4+) naturally use it without forcing/reminders. Now deemphasized—nice for visibility but not necessary.
Models eat harness for breakfast: remove prompting interventions. Every launch, review system prompt, cut unneeded reminders.
New models unlock features. Code review: Tried multiple times, launched simple /codereview. Recent models (Opus 45/46, Sonnet 4.6) enable reliable version—run multiple agents to traverse codebase, synthesize real issues. Engineering team relies on it pre-merge; catches majority of bugs.
Build at edge of possible (next 6 months), prototype identifies gaps, swap in new model.
Think in building blocks:
- Individual task success: clear prompt → consistent acceptable output (merge/share).
- Higher success → multiple tasks (multi-coding surged end of 2025, increasing).
- Next: 50–hundreds of Claudes simultaneously. Infrastructure for remote runs (RAM limits), interfaces to track tasks, agent self-verification for trust, self-improving via feedback incorporation.
AI gives leverage. Automate repetitive manual tasks with Claude Code/Co-work. Offload tedious parts you hate; AI learns from your examples, generalizes. Focus on creative parts, do more: pet projects, untackled ideas. Gain 20% extra time.
Solve problems for yourself: repetitive tasks, floating ideas.
Push automations to 100% success (not 95%—not true automation). Last 5–10% takes time; building slower initially. Invest elbow grease, feedback to teach preferences.
Example: Cat teaching Co-work inbox zero for Gmail—time-consuming, not there yet. Podcast host's spam email filter: 95%, misses some—push to perfect.
Anthropic improving customization flow (skills, feedback, updates)—currently requires too many concepts.
Build daily-used apps for real value/leverage, not one-shot prototypes.
Avoid over-customizing (adding endless skills/MCPs)—distracts from launching products/features.
There's fun in customizing products and making them hackable to work well for individual needs, but there's a limit to its usefulness. Some people spend excessive time customizing, neglecting sleep and core tasks. Simple setups often work better. People showcase overly optimized setups on Twitter, prioritizing the setup over actual building.
Andre Karpathy recently tweeted about the divide between early ChatGPT users who dismissed it as terrible and gave up on AI, remaining cynical, and those using it for coding who recognize its full power. Both sides fail to understand each other. Advice: use AI for real tasks to see its improvements.
2024 products were chat-based; Claude Code generation is action-based. The aha moment is when Claude acts on behalf of users, doing tasks itself rather than just advising. Watching the Claude Chrome extension fill out forms demonstrates this capability.
How Asia Works: Story of economic development, policies, and governments creating long-lasting successful economies.
The Technology Trap: Examines past technology revolutions like the Industrial Revolution and computer revolution, their effects on workers, and lessons to ensure smooth transitions.
Paper Menagerie: Book of short stories on coming of age, AI, and self-discovery.
Drive to Survive: Satisfying portrayal of obsession with a singular engineering goal and purity of pursuit.
Free Solo: Documentary on Alex Honnold climbing El Capitan without a harness, showcasing pure achievement, mental focus, and high stakes where a single mistake means death. Cat Wu is a rock climber; the more she knows, the more impressive it becomes—even moves she'd never do one foot off the ground with a rope in a gym. Mentioned the sad documentary on the younger climber on ice mountain.
Waymo has most changed life outside Claude products. Used twice daily for commuting. Benefits:
- No guilt if Waymo waits; less pressure to be curbside exactly on time.
- Enables productivity: can join work calls without rudeness concerns (no human driver overhearing, no music adjustments). Gains back 30 minutes daily.
Happy to pay 2x premium over Uber/Lyft. Changes vernacular at Anthropic: people say "Is the Waymo here?" instead of generic ride-share.
"Just do things." With first principles thinking and clear optimization goals, deduce the right action, articulate to stakeholders, and execute. Jobs are fake: understand constraints, act quickly, learn from mistakes, apologize or fix.
Empowers crossing team boundaries, rigid roles, and codebase scopes. Fosters agency and bias toward action. Gained from startup experience at Scale (20 people), no processes, big problems, full empowerment across roles (sales, ops, engineering) to solve ambitious issues.
School/college teaches following instructions for grades; unlearn to do what's needed, even if others think it's dumb.
Manifesting—also her favorite sticker.
AGI diffusion across society will take time; immediate focus is helping bring the world along. Non-serious: move to Fontainebleau, live among 10,000 boulders, rock climb extensively. Read 1-2 books/week (current rate 0.5); backlog huge. Learn from history and unexplored topics like physics, robotics, hardware, aerospace—excited even if AI knows it.
@catwoo on Twitter; tag or DM (reads all, responds to some). Most helpful: share edge cases, errors, reproducible failures in Claude Code or Cowork for model/harness improvements. Team active on Twitter, channels user love/successes and issues for broader action. Encourages sharing problems.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
